HPL-2002-42 - Speech-As-Data Technologies for Personal Information Devices

8/12/2019 HPL-2002-42 - Speech-As-Data Technologies for Personal Information Devices

1/10

Speech-As-Data Technologies for

Personal Information Devices

Roger C.F. Tucker, Marianne Hickey, Nick Haddock1

Client and Media Systems LaboratoryHP Laboratories BristolHPL-2002-42March 18th, 2002*

E-mail: [email protected]. [email protected], [email protected]

speech-as-data, speechrecognition,wordspotting,speechcompression,audiosummarisation

For small, portable devices, speech input has theadvantages of low-cost and small hardware, can be used onthe move or whilst the eyes & hands are busy, and isnatural and quick. Rather than rely on imperfect speechrecognition we propose that information entered as speechis kept as speech and suitable tools are provided to allowquick and easy access to the speech-as-data records. Thispaper summarises the work carried out at HP Labs and itspartners at Cambridge University on the technologies

needed for these tools - for organising, browsing, searchingand compressing the stored speech. These technologies goa long way towards giving stored speech thecharacteristics of text without the associated inputproblems.

*Internal Accession Date Only Approved for External Publication1 Consultant working with HP

Copyright Hewlett-Packard Company 2002


2/10


3/10

3

2. Is this information better expressed in speech or text? Personal feelings, perhaps even spoken by

the person whose entry it is, are more easily conveyed using speech.

3. Is this information that would take a long time to enter by hand e.g. an address or note? Speech

saves time now at the expense of possibly more time later.

4. Is this information that I may not need, but would like to keep just in case? Fax numbers,

addresses and titles are fields in this category. In this case it is easy to read them in and then they

are there if ever needed.

Figure (1): Speech-enabled Contacts Application

2.2 Eyes-busy entry of information

To benefit from the portability of the device, it must be possible to fill in these fields in an eyes-free way.Weve done this by recording a voice memo structured with keyword labels [4]. The keywords can be

pre-defined or user defined, but in either case they mark the start of a section of speech that should be cut

out and inserted into a particular field. For example, whilst driving you may see something on the side ofa truck and record:

"Contacts entry Work Number is 408 927 6353 Name is Hamilton & Co Removals

Comment is the truck says something about extra heavy items a specialty, could get that pianoshifted at last"

The recording is then processed in three stages:

1. The very first keyword is found, and used to file the recording into the right application.

2. Subsequent keywords are found and used to segment the recording, each segment being sent tothe appropriate field in the desired application.

3. Where possible, the speech data is recognised and entered as text.

Since the entry name is not recognised, the entry will not be alphabetically placed, but just sit at the top ofthe Contacts list until at least the name is transcribed. To allow the detection of keywords to be done

reliably and with a simple recogniser, the user is required to pause before (and possibly after) eachkeyword phrase. By requiring the pause to be a long one, e.g. more than one second, non-keyword

phrases can be rejected quite easily as long as the pause detection is robust.

There isspeech

data for

this field

There isspeech data

for this record

There is speechdata for the

Work Tel field

Play speech

associated with

highlightedfield

Record

speech for

highlightedfield

Delete speech

associated withhighlighted

field


4/10

4

3. Browsing Speech Data

3.1 Graphical Interface

Whilst ideally all speech should be recorded in small easily accessible chunks, this will not always bepossible - not all information to be recorded will fit neatly into the available fields of an application. Once

more than 10 seconds or so of speech is recorded in one go, navigating becomes an important issue. Asimple but effective solution is to provide a 2-D random-access interface [5]. Figure 2 shows such an

interface developed at HP Labs [6], where a 100 second recording is displayed as a series of lines about 8seconds long.

Figure (2): Random-Access Interface for Longer Speech Records

The speech record is divided into chunks based on the pauses in the recording. Whilst ideally the chunks

would match phrases in the speech, it is equally important that they are a suitable length for skipping and

marking. The algorithm creating the chunks [7] has the concept of a target length (about 5 seconds),

which it tries to achieve within the constraints of not breaking up a continuous phrase. The ability to marka place in the speech is very important, and used properly ensures that most of the speech need only be

listened to once. A number of different markers could be made available for different categories - we

have used icons for phone numbers and dates/times to good effect.

3.2 Textual Summaries

So far we have not considered the possible role of large vocabulary speech recognition if it were

available, either on the client device or on a network server. This allows complete speech records to be

transcribed automatically, but with a substantial percentage of errors especially for unconstrained speech.

One approach is to leave these for the user to cope with, as with the AT&T SCAN interface for speech

records [8]. Our approach is topartiallytranscribe the speech. The aim is to give enough of a textual

summary to enable the user to quickly get to the part of the speech data they need. The user can thenlisten to get the full information and if they want to, complete or correct the transcription process.

Two components are used to provide the summary - an acoustic confidence score for each word, and aninverse frequency score. Other features such as prosody [9] can also be used. The acoustic confidencescore can be derived a number of ways, though in the case of the ANN/HMM hybrid recogniser [10] we

used, the confidence scores were obtained directly from the phone probabilities [11]. The inverse

frequency score is the number of times a word appears in a recording divided by the number of times itoccurred in the language model and normalised by the length of the recording. The two scores arecombined:

*inverse frequency +*acoustic confidence (1)

where the relative weighting factors and depend on the importance of accuracy versus coverage.

Playback speed

50-200%Back onechunk

Forward

one chunkMark a

chunk while

it's playing


5/10

5

The summary is created by first automatically transcribing the recording using a large vocabulary

recogniser, and then scoring each word according to Equation 1. Then, an N-word window (typically 10-30 words long) is slid over the transcript, and a score for each N-gram calculated. A single parameter,

number of words per minute, determines both the choice of N and how many N-grams are displayed.Table 1 shows some partial transcriptions of a one-minute section of the May 22, 1996 broadcast of CNN

EPN. Full details of the system are given in [12] along with evaluation results.

Size Summary of one minute of speech

2 keywords

per minute

FERRY, OVERCROWDED

1 10-gram per

minute

FERRY SINKING IN WHICH HUNDREDS OF PASSENGERS DIED [SIL] THE

FERRY

1 30-gram per

minute

INTO YESTERDAY'S FERRY SINKING IN WHICH HUNDREDS OF

PASSENGERS DIED [SIL] THE FERRY WENT DOWN ON LAKE VICTORIAWITHIN SIGHT OF SHORE [SIL] THE GOVERNMENT SAYS FIVE HUNDRED

FORTY NINE PEOPLE ON

20 (potentiallyoverlapping)

10-grams perminute

: [SIL][SIL] AFTER A FOUR-DAY HEARING IN FEDERAL COURT A JUDGEHAS STRIPPED YESTERDAY'S FERRY SINKING IN WHICH HUNDREDS OF

PASSENGERS DIED [SIL] THE FERRY WENT DOWN ON LAKE VICTORIAWITHIN SIGHT THIRD CLASS PASSENGERS [SIL] CRUISE TICKETS ONLYGAVE THEM A PLACE TO SIT OR STAND ON DECK [SIL] THOSE PASSENGERS [SIL] BUT HUNDREDS MORE IGNORED WARNINGS AND

JAMMED ABOARD THE VESSEL [SIL] TO MAKE THE TRIP [SIL] FUMBLE

Table (1): Example Partial Transcriptions

4. Searching Speech Data

4.1 Keyword Searching

Because browsing speech is not as fast as text, a good search facility is even more important than for text.

A search facility that uses spoken queries also provides a generic eyes-free way of accessing any of thedata on the device. Searches can specify a field e.g. search the "File As" field in Contacts for a specificname, or can be a free search through all the stored speech in the system. Matching one isolated utterance

against another (as in the first case) is quite fast and accurate using conventional isolated word

recognition techniques, but spotting for a phrase embedded in audio is both slow and difficult since theendpoints are not known. Whilst there is a large body of research on wordspotting, our task is different to

a conventional wordspotting application. Normally, a single utterance is searched for a number of words

for which good models exist, and the task is primarily to determine which word was spoken. In our case

we are searching a potentially large number of utterances for one word that we have just one spokenexample of, and the task is to decide if it is there or not. The two technical issues raised by these

differences are how to use the spoken keyword in the wordspotter, and how to make the search go fastenough.

The Video Mail Retrieval (VMR1) database [13] is a database of 35 speakers recorded specifically for

testing this kind of wordspotting system. As well as a number of read-speech training sentences, each

speaker recorded 20 spontaneous speech messages and then 35 isolated keywords frequently occurring inthe messages. Performance is measured in terms of the percentage of hits scoring higher than the Nth

false alarm (or percentage misses scoring lower). Exactly which value of N is most meaningful is quitedifficult to decide as it depends on how fast the hits can be assessed by the user, how much audio there is,and how many false alarms they would be prepared to listen to before finding the right one. To get around

this difficulty of choosing an operating point, a metric known as the Figure of Merit (FoM) [14] can be

used, which in essence gives the percentage hits averaged over 1 to 10 false alarms per hour. To give anidea of the performance of state-of-the-art speech recognition on this task, Table 2 shows the FoM for the

different keywords in the VMR database, using the Abbot recognition system [10] as the wordspotter. Forthese results, the keywords have been entered as text and the phonetic transcription looked up in a


6/10

6

dictionary. Whereas the longer more complex words can be found quite reliably, the short words are

easily missed.

FoM Range Words (lowest FoM first)

Below 30% WORD

30-40% MAIL, DATE, VIDEO, SCORE, OUTPUT

40-50% FIND, BADGE, ACTIVE

50-60% TIME

60-70% KEYWORD, STAFF, SENSOR, SPOTTING, SEARCH

70-80% RANK, MESSAGE, INDIGO, DOCUMENT, PLAN, DISPLAY

80-90% PROJECT, ASSESS, MEETING, PANDORA,

90-100% WORKSTATION, MANAGE, MICROPHONE, INTERFACE,

NETWORK, CAMERA

Table (2) Figure of Merit Scores for Keywords in VMR database

The next two sections address the two key issues for wordspotting in the context of audio search: how to

model the spoken keyword, and the speed of the search.

4.2 Modeling the Keyword

The problem here is that the search query will have the word spoken in isolation and probablyemphasized, but the utterance with it in will have it embedded in continuous speech. We t herefore cannot

use the query speech directly, but instead need to pass it through a phoneme recogniser to convert it to a

string of phonemes. The keyword for the search is then modeled by concatenating the appropriatephoneme models.

In [15], Knill and Young investigate a number of enhancements to this basic approach, two of which give

a small improvement. The first enhancement is to use the N-best phoneme strings from the phoneme

recogniser in the search. Since this tends to increase the false alarms as well as the true hits, it only

improves performance in the cases where the true phoneme string wasn't the most probable. Around 3%overall reduction in the number of missed words is achieved with N=7. This rather small improvement isat the expense of a slower search, as there are now N versions of the keyword to look for. The second

enhancement is to use two examples of the keyword rather than one. If the two phoneme strings generated

are different, both can be used, though it is as good just to use the one which gives the best Viterbi scorewhen force-matched against the other spoken example. This achieves around 10% reduction in thenumber of missed words. Disappointingly, even the best spoken keyword system increased the number of

missed words by a factor of at least 30% compared to using a dictionary-derived phoneme string for the

search word. Nevertheless, a quick search and good user interface can make up for some lack ofperformance, and the user can easily improve the search by using longer words or even phrases.

4.3 Speed of Search

We would ideally like the search to go fast enough to appear instantaneous. But unlike a textual search,

where all the results can be displayed and assimilated by the user very quickly, with audio each hit has to

be listened to. Even if 5 seconds of speech is enough context to check a potential hit, it will take the user25 seconds to check the top 5 hits. This time can be incorporated into the search by playing the mostprobable hit found so far whilst the search is taking place. As an example let us assume that there is a

total of 30 minutes of speech to be searched and that in 5 seconds 7.5 minutes of speech can be searched.

The hits delivered to the user could be:

1. after 5 seconds- best hit in the first 7.5 minutes

2. after 10 seconds- best hit in the first 15 minutes not already played

3. after 15 seconds- best hit in the first 22.5 minutes not already played

4. after 20 seconds- best hit in whole 30 minutes not already played


7/10

7

5. after 25 seconds next best hit in whole 30 minutes not already played

As the first three hits are from partial searches, this strategy is not perfect, but it requires 90 times real-time rather than the 360 times real-time needed to do the whole 30 minutes in the first 5 seconds.

To achieve such a fast search on the modest computing platform we expect to find in a portable device,the speech has to be pre-processed when, or soon after, it is recorded. Fortunately, the structure of a

wordspotting system lends itself to this.

A standard wordspotting system performs two recognition passes - one with keywords and fillerstogether, the other with just the fillers. The second pass allows the keyword scores to be normalised

against speaker and acoustic variations. The fillers can simply be phoneme models, and in our case are the

same models used to transcribe the spoken keyword. There is a large amount of duplication of effort inthe two passes, and the key to doing a fast search is to do the filler-only pass (half of the computation)and as much as possible of the keyword-and-filler pass ahead of time. This must not be at the expense of a

large amount of storage as in the lattice-based approach of [16]. In [17], Knill and Young investigate

various techniques for this, the two most effective being:

Approximated keyword recognition pass. After the filler-only pass, the resulting phoneme

boundaries and Viterbi scores at each boundary are stored. The keyword-and-filler pass

performed at search time is then restricted to start and finish on a stored phoneme boundary

rather than at arbitrary start and end frames. This simplification leads to a further 80% reductionin search time without affecting performance in any significant way.

Pattern Matching of Phoneme Strings. A string-match can be performed between the keywordand the filler-only pass output extremely quickly, and whilst this is not an accurate enough

process for doing the search, it can be used to eliminate a lot of the search space. Since both

phoneme strings are output from a recogniser, the match must allow for recognition errors. Thisis done by having penalties for substitution, deletion and insertion, and using a DynamicProgramming (DP) algorithm to perform the match. Preprocessing in this way eliminates 85% of

the search space with only a small drop in performance.

These two techniques give a 33 fold speed up, so the basic wordspotter only needs to work a little fasterthan real time to allow a reasonable search.

In the case of a hybrid ANN/HMM recognition system such as Abbot [10], there is a further opportunityto speed up the search by pre-computing and storing the posterior phone probabilities for each frame. Thisleaves very little computation to be done at search time. On a PC with 500MHz dual-processor and

512MB RAM we found we could search for 10 keywords (e.g. the 10-best transcriptions of a spoken

keyword) at 300 times real time. This was withoutmaking use of either of the two speed-up techniquesmentioned above. Since there are 45 probabilities per 16ms frame (for UK English), some compression is

needed to reduce the storage requirements, which even with each probability represented by just one byte

would be 1Mbyte for every 6 minutes of speech. We were interested to see if this could be reduced to alevel which required little extra storage compared to even the lowest rate speech coders, and so targeted

just 10 bits per 45 probabilities, i.e. 625 bits/sec. We achieved this using vector quantisation and a

log(1+100x) transform applied to the probabilities before computing the distance measure, resulting in anerror rate within 10% of the unquantised baseline system [18].

5. Compression of Speech Data

5.1 Requirements

For speech-as-data to be viable on a small device, the speech must not take up a lot of memory. There are

numerous speech coding schemes, many being published standards, optimised for telephony. These offer

a high perceptual quality (i.e. speech must sound natural and like the talker), but can afford to be a little

unintelligible as the hearer can always request a repeat. For this reason a restricted 3.4kHz bandwidth has

been used for telephony from the early days of frequency-division multiplexing until now.

For speech-as-data the perceptual quality need not be high, but the speech must be as intelligibile as

possible. Phone numbers and alphanumeric details must be understood without error otherwise the entireconcept fails. At the same time a high degree of compression is needed. The military solution to a similar

requirement (high intelligibility, low bit rate) is the LPC10 vocoder [19]. This is a speech coder that reliescompletely on encoding the parameters of a model of speech production. Advances in parametric coding


8/10

8

[20] have substantially improved the quality (perceptual and intelligibility) of the LPC vocoder, and it is

now close to the quality needed for speech-as-data whilst keeping a very low bit rate.

5.2 A Wideband Low-Bit-Rate Speech Coder

In [21], a speech coder is described based on LPC vocoding that improves intelligibility by extending thebandwidth of the encoded speech to 8kHz. The input signal is split into two bands, and the 0-4kHz band

is encoded with a 10thorder LPC model, and the 4-8kHz band is encoded with a 2

ndorder LPC model.

The parameters are put through predictors and the residual is encoded using variable-rate Rice coding.Variable-rate coding is a luxury that stored speech can benefit from, which transmitted speech cannot. For

extra intelligibility, the frame length is reduced from 22.5ms to 16ms. This reduction in frame length does

not increase the bit rate proportionately as the smaller frame size is more predictable. At 2.4kbit/sec, thehigh-band parameters take up about 500 bit/sec. The interesting result from the Dynamic Rhyme Test

(DRT [22]) reported in [21] is that when comparing the wideband and narrowband versions with both

optimised for 2.4kbit/sec operation, the wideband coder performs 2.4 points on the DRT scale better thanthe narrowband coder. This demonstrates that even at this very low bit rate, bandwidth is more useful than

finer quantisation.

5.3 Recognition of Compressed Speech

Various studies have shown that recognition performance degrades when recognising from speechcompressed below 16kbit/sec [23,24]. For very low bit rate coders the problem is even worse. The

ANN/HMM recogniser keyword-searching system described in section 4.3 can bypass this problem by

preprocessing the speech at record time and therefore before compression, and can store the resulting datain less than 1kbit/sec. But for a general HMM-based recogniser, record-time processing can only go as far

as encoding the acoustic features, which still require at least 4kbit/sec storage [25,26,27]. These could be

stored alongside the speech, more than doubling the storage requirements, but in fact the tworepresentations (coded speech and acoustic features) are only slightly different forms of the same

information. It should be possible to derive the acoustic featuresfromthe coded speech, without spending

the time decoding and re-analysing the speech with the associated loss in quality of a 'tandem' encoding.

The process of deriving spectrum-based features from the LPC parameters stored in a speech coder and

the recognition performance of such a system are described in [18]. In fact the performance is as good asone with features derived directly from the original speech. However, the two parameter sets are clearly

different since the error rate doubles when training is performed with one parameter set and testing with

the other. Frame size appears to be an issue though - matching the coder frame size to the recogniser's16ms frame gives virtually half the error of matching the recogniser's frame size to the coder's 22.5msframe (using LPC-derived features for training and testing in both cases). This result may well be

influenced by the fact the recogniser is designed to operate at 16ms. But it does mean that you areunlikely to get optimal performance just by using the LPC parameters from a standard coder, as framesizes are normally in the 20 to 30ms range.

The performance is dependent on the amount of quantisation. Using the quantisation scheme described in

[21] at 2.4 kbit/sec increases the error rate from 5.5% to 7.6%. The error rate is reduced to 7.1% if thetraining is performed on unquantised parameters, even though the recognition has to be done on quantised

parameters, since the quantisation affects the variance but not the means of the parameters. This increasein error can be eliminated with finer quantisation at the expense of using a higher bit rate than 2.4

kbit/sec.

5.4 Reconstruction of Speech from Acoustic Features

The logical alternative to deriving acoustic features from coded speech is to reconstruct speech from the

acoustic features derived for speech recognition. This is the approach taken in [28]. By adding a voicing

and pitch detector to the acoustic feature extraction, speech similar in quality to LPC vocoding can bereconstructed. This approach has the advantage that a standard speech recogniser can be used (no

retraining needed), but the disadvantage that the speech quality is not quite as good and the bit-rate ishigher.


9/10

9

6. Conclusion

In this paper we have described technologies that enable speech to be used effectively as a data type in

small, portable devices by providing means for organisation, browsing, searching and compression of the

speech.

If the speech can be broken down into short segments only a few seconds long, and associated with the

fields of an application, it will be easy to find the information and quick to listen to it. This organisation

can be done either manually, by allowing the user to enter speech into any field as an alternative to text,or automatically with simple speech recognition responding to keyword labels that the user inserts intothe recordings. Whilst the manual approach is preferable, the use of keyword labels allows information to

be entered on-the-move, with eyes and hands busy.

For browsing longer recordings that cannot be broken up into short segments, a 2D random-accessinterface should be used in preference to tape-recorder style controls. This allows easy skimming and

replay, and the ability to mark important sections with a visual icon for future reference.

When a large vocabulary recogniser is available as when connected to a larger machine, its output can beused without the need for checking and correction by applying summarisation techniques based on

confidence score and inverse word frequency. The text becomes a quick way of locating relevant spoken

information and need never be corrected unless the information is needed in textual form.

The other way of locating information is through a keyword search, where the user speaks a word or

phrase, which locates the speech they are looking for. This search is more effective the longer thekeyword. Ironically, searches with spoken keywords perform less well than searches with the samekeyword entered as text, as the spoken keyword cannot be used in the search directly, but has to be

transcribed into phonemes first. Speaking the word twice helps. A big issue is search speed - something inthe order of 100 times real time is required, within the limited capabilities of a small device. This can be

achieved by a combination of preprocesing at record time and simplifications to the search space. If an

ANN/HMM hybrid recogniser is used, the posterior phone probabilities can be precomputed and stored inunder 1kbit/sec, allowing a very fast search indeed.

To avoid using up memory with a lot of stored speech, a good compression scheme is needed. It needs to

have high intelligibility to allow names, numbers and addresses to be retrieved from the speech - it also

needs to work well with speech recognisers. This can be achieved with wideband (8kHz) encoding, whichrequires only 500 bit/sec extra and even at rates as low as 2.4 kbit/sec gives greater intelligibility than

narrowband encoding. Good recognition performance can be achieved by deriving the acoustic features

directly from the LPC parameters of the speech coder, but for best performance the LPC frame rate needs

to be reduced to 16ms.

These techniques together go a long way towards giving stored speech the characteristics of text. Thisenables speech to be used for information input without the need for checking and correcting speech-to-

text output, and with very modest computational resources. As a result, the device can be used on-the-

move with eyes and hands busy, and can be small, low power and low cost without compromisingfunctionality.


10/10

10

References1. Pearce, D. Enabling new speech driven services for mobile devices: An overview of the ETSI standards

activities for distributed speech recognition front-ends. In: Proc American Voice I/O society (AVIOS), May

22-24, 2000, San-Jose CA, USA.

2. Stifleman L J, Arons B, Schmandt C et al. VoiceNotes: A Speech Interface for a Hand-Held Voice

Notetaker.In: Proceedings of INTERCHI, April 1993, ACM Press, 1993, pp 179-1863. Arons B. Interactively Skimming Recorded Speech. PhD Thesis, M.I.T., 1994

4. Haddock, N. Structuring Voice Records Using Keyword Labels. In: Proceedings of CHI '96, ACM Press,

19965. Ades S, Swinehart DC. Voice Annotation and Editing in a Workstation Environment. In: Proceedings of

AVIOS '86, 1986

6. Loughran S, Haddock N, Tucker R. Voicemail Visualisation from Prototype to Plug In: TechnologyTransfer through Component Software, HP Labs Technical Report HPL-96-145 1996.

7. Tucker RCF, Collins MJ. Speech Segmentation. U.S. Patent 6,055,495. April 2000.8. Whittaker S, Hirschberg J, Choi J, Hindle D, Pereira F, Singhal A. SCAN: designing and evaluating user

interfaces to support retrieval from speech archives. In: Proceedings of SIGIR99 Conference on Research

and Development in Information Retrieval.1999.pp 26-33.

9. Koumpis K, Renals S. The Role of Prosody in a Voicemail Summarization System. In:ISCA Workshop onProsody in Speech Recognition and Understanding, Red Bank, NJ, USA, Oct. 2001.

10. Robinson AJ, Hochberg MM, Renals S. The Use of Recurrent Neural Networks in Continuous Speech

Recognition. In: Automatic Speech and Speaker Recognition. Kluwer, 1995, Chapter 19

11. Williams G, Renals S. Confidence Measures for Hybrid HMM/ANN Speech Recognition. In: Proceedings ofEurospeech '97, 1997

12. Valenza R, Robinson T, Hickey M et al. Summarisation of Spoken Audio through Information Extraction.In: Proceedings of ESCA ETRW Workshop on Accessing Information in Spoken Audio, April 1999, pp111-116

13. Jones GJF, Foote JT, Sparck Jones K et al. Video Mail Retrieval using Voice: Report on Keyword

Definition and Data Collection. University of Cambridge Computer Laboratory, Tech. Report No. 335, May,

199414. NIST. The Road Rally Word-Spotting Corpora (RDRALLY1). NIST Speech Disc 6-1.1, September 1991.

15. Knill KM, Young SJ. Techniques for Automatically Transcribing Unknown Keywords for Open Keyword

Set HMM-Based Word-Spotting. Cambridge University Eng. Dept. Technical Report CUED/F-INFENG/TR230, Sept 1995

16. James DA, Young SJ. A Fast Lattice-Based Approach to Vocabulary Independent Wordspotting. In:Proceedings ICASSP'94, Adelaide 1994, IEEE.

17. Knill KM, Young SJ. Low-cost Implementation of Open Set Keyword Spotting. In: Computer Speech and

language (1999) 13, 243-266

18. Tucker R, Robinson AJ, Christie J et al. Recognition-Compatible Speech Compression for Stored Speech.

In: Proceedings of ESCA ETRW Workshop on Accessing Information in Spoken Audio, April 1999, pp 69-

7219. Tremain TE. The Government Standard linear predictive Coding Algorithm: LPC10. Speech Technology

1982, pp 40-49

20. McCree AV, Barnwell III TP. A Mixed Excitation LPC Vocoder Model for Low Bit Rate Speech Coding.

IEEE Trans. Speech and Audio processing, July 1995, 3:242-25021. Seymour CW, Robinson AJ. A Low-Bit-Rate Speech Coder using Adaptive Line Spectral Frequency

Prediction. In: Proceedings of Eurospeech '97, 1997

22. Voiers WD. Diagnostic Evaluation of Speech Intelligibility. In: Speech Intelligibility and Speaker

Recognition, Mones E Hawley, Ed., Dowden, Hutchinson & Ross, Inc., 1977, pp 374-387

23. Euler S, Zinke J. The Influence of Speech Coding Algorithms on Automatic Speech Recognition. In:Proceedings ICASSP'94, Adelaide, IEEE, 1994, I:621-624

24. Lilly BT, Paliwal KK. Effect of Speech Coders on Speech Recognition Performance. In: Proceedings

ICSLP'96, pp 2344-2347

25. Ramaswany GN, Gopalakrishnan PS. Compression of Acoustic Features for Speech Recognition in Network

Environments. In: Proceedings of ICASSP'98, IEEE, 1998 2:977-98026. Digalakis V, Neumeyer LG, Perakakis M. Quantization of Cepstral Parameters for Speech recognition over

the WWW. In: Proceedings of ICASSP'98, IEEE, 1998 2:989-99227. Speech Processing, Transmission and Quality aspects (STQ); Distributed speech recognition; Front-end

feature extraction algorithm; Compression algorithms. ETSI ES 201 108 version 1.1.2. April 2000.

28. Chazan D, Hoory R, Cohen G, Zibulski M. Speech Reconstruction from MEL Frequency CepstralCoefficients and Pitch Frequency, In: IEEE ICASSP 2000, Vol 3, pp 1299-1302

HPL-2002-42 - Speech-As-Data Technologies for Personal Information Devices

Documents

Transcript of HPL-2002-42 - Speech-As-Data Technologies for Personal Information Devices